Re: [PATCH] hrtimer: Replace cpu_base->active_bases with a direct check of the active list

[Ingo Molnar]

* Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:

> On Thu, Apr 09, 2015 at 09:09:17AM +0200, Ingo Molnar wrote:
> > 
> > * Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > 
> > > On Thu, Apr 09, 2015 at 08:28:41AM +0200, Ingo Molnar wrote:
> > > > Btw., does cpu_base->active_bases even make sense? hrtimer bases are 
> > > > fundamentally percpu, and to check whether there are any pending 
> > > > timers is a very simple check:
> > > > 
> > > > 	base->active->next != NULL
> > > > 
> > > 
> > > Yeah, that's 3 pointer dereferences from cpu_base, iow you traded a 
> > > single bit test on an already loaded word for 3 potential cacheline 
> > > misses.
> > 
> > But the clock bases are not aligned to cachelines, and we have 4 of 
> > them. So in practice when we access one, we'll load the next one 
> > anyway.
> 
> $ pahole -C hrtimer_clock_base defconfig-build/kernel/time/timer.o 
> struct hrtimer_clock_base {
>         struct hrtimer_cpu_base *  cpu_base;             /*     0     8 */
>         int                        index;                /*     8     4 */
>         clockid_t                  clockid;              /*    12     4 */
>         struct timerqueue_head     active;               /*    16    16 */
>         ktime_t                    resolution;           /*    32     8 */
>         ktime_t                    (*get_time)(void);    /*    40     8 */
>         ktime_t                    softirq_time;         /*    48     8 */
>         ktime_t                    offset;               /*    56     8 */
>         /* --- cacheline 1 boundary (64 bytes) --- */
> 
>         /* size: 64, cachelines: 1, members: 8 */
> };
> 
> They _should_ be aligned :-)

Maybe, but they aren't currently - and aligning them has costs as 
well.

> > Furthermore the simplification is measurable, and a fair bit of it 
> > is in various fast paths. I'd rather trade a bit of a cacheline 
> > footprint for less overall complexity and faster code.
> 
> cacheline misses hurt a lot, and the bitmask isn't really complex.

See my other mail: in practice we already dirty all of these 
cachelines in the hrtimer irq...

Thanks,

	Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/